摘要
针对传统基于内容的纹理图像检索方法中抽取固定特征和相似性度量存在累积误差的不足,提出一种新的基于学习的Log-gabor子空间特征优化的旋转不变纹理图像检索方法.首先根据Log-gabor分解的幅度和相位构造旋转不变多尺度广义粗糙度纹理描述子;然后一方面通过支持向量机粗分类器缩小被检索图像的分类范围,另一方面通过有监督训练构造支持向量特征选择器,选择优化的自适应纹理描述特征作为进一步检索的输入;在相似性度量过程中提出特征量化消除累积误差的影响.仿真实验结果表明,算法对任意角度纹理图像的检索都具有较好的鲁棒性.
In order to overcome the defects of general image retrieval that using same texture feature vector and existing accumulative error in similarity measure, a novel Rotation Invariant Texture Retrieval Method Based on Learning Log-Gabor Subspace Feature Optimization was proposed. The rotation-invariant vector is formed according to the amplitude spectrum and frequency spectrum of the Log-Gabor decomposition. Support Vector Machine (SVM) is introduced as a coarse classifier to reduce the ranges of texture image distribution. Supervised training procedure is adpoted to obtain the optimization discriminative feature subspace in each type for following step. A quantization operator is introduced to get ride of the affection of accumulative error in the last step of similarity measure. Experiments proved the efficiency of this method for the general texture images.
出处
《哈尔滨工业大学学报》
EI
CAS
CSCD
北大核心
2006年第5期669-672,共4页
Journal of Harbin Institute of Technology
基金
国家自然科学基金资助项目(60272073)
关键词
纹理图像检索
机器学习
相似性度量
子空间特征选择
texture image retrieval
machine learning
similarity measure
subspace feature selection